CN102455786B - System and method for optimizing Chinese sentence input method - Google Patents

Abstract

The invention discloses a system and a method for optimizing a Chinese sentence input method. The method comprises the following steps of: after obtaining a plurality of candidate Chinese sentences, calculating the characteristic vectors of selected characteristics for each Chinese sentence, thus obtaining the characteristic vectors reflecting the language habit, the grammar and the semantic meaning of the sentence; next, performing inner product operation on the characteristic vectors and a trained characteristic weight to obtain the score of each Chinese sentence; and after ordering the plurality of candidate Chinese sentences according to the score of each Chinese sentence, displaying the ordered candidate Chinese sentences through a candidate Chinese sentence list display module. The candidate Chinese sentence list obtained according to the ordering method provided in the method for optimizing the Chinese sentence input method is ordered so as to more accord with the language habit of a user and the grammar and semantic meaning characteristics, so that the candidate Chinese sentence list is optimized and the accuracy of the candidate Chinese sentence list is improved.

Description

A kind of optimization system of Chinese sentence input method and method

Technical field

The present invention relates to the input in Chinese technology of computing machine, particularly a kind of optimization system of Chinese sentence input method and method.

Background technology

At present, carry out input in Chinese at computing machine, occurred the input of Chinese sentence.The input of Chinese sentence is exactly the pinyin string that computing machine receives user's input, and this pinyin string is translated as to corresponding Chinese sentence.

Fig. 1 is the system of the Chinese sentence input method of prior art, comprising: input method engine modules, identification translation module, candidate's Chinese sentence list display module, selection module and output module, wherein,

Input method engine modules, resolves the pinyin string that obtains user's input for the input method according to setting, send to identification translation module;

Identification translation module, for the pinyin string to the input of input method engine modules, identify after translation, obtain multiple candidates' Chinese sentence, according to certain strategy, such as according to Chinese language model scoring, sort, obtain corresponding candidate's Chinese sentence list, send to candidate's Chinese sentence list display module;

Candidate's Chinese sentence list display module, for the list of show candidate Chinese sentence, by selecting module controls to select to obtain a Chinese sentence, exports by output module;

Select module, meet for of selecting candidate's Chinese sentence list display module the Chinese sentence that user requires most.

In this system, identification translation module is vital, and in fact its effect is exactly the transfer process of pinyin string to Chinese sentence, can adopt noisy channel model to be described:

$\hat{H}=\underset{H}{\arg \max }P\left(H\right|Y)=\underset{H}{\arg \max }\frac{P\left(Y\right|H)P\left(H\right)}{P\left(Y\right)}\≈\underset{H}{\arg \max }P\left(Y\right|H)P\left(H\right)$ Formula (1)

In formula (1), Y represents Chinese phonetic alphabet string, and H represents Chinese character string.Conventionally for convenient, claim that P (Y|H) is sound-word transformation model, claim the language model that P (H) is Chinese.

In the input process of Chinese sentence, due to phonetically similar word in Chinese and polyphone a lot, and the pinyin string voiceless sound of conventionally input adjusts, the corresponding a lot of Chinese characters of phonetic, so a word corresponding to pinyin string just can be combined into the Chinese sentence of a lot of candidates.In order to select candidate's Chinese sentence list that possibility is the highest from countless candidate's Chinese sentences, just need search procedure, because Chinese character group word is flexible, search volume is very large, in order to meet the requirement of user to the response time, constantly beta pruning in search procedure, finally can only provide candidate's Chinese sentence list that possibility is the highest and select for user, the list of common this candidate's Chinese sentence is n-best, represents front n best candidate's Chinese sentence.

Although the system of existing Chinese sentence input method has adopted some technological means in the process that generates the list of candidate's Chinese sentence, such as adopting formula (1) according to the Chinese language model arranging, candidate's Chinese sentence to be translated, but also have following problem:

1) for some factors or feature, such as: language model, the sequence of candidate's Chinese sentence is had a significant impact, but owing to being subject to response time or technical limitation, more the language model of high-order can not be used for candidate's Chinese sentence to sort;

2) Chinese language model that identification translation module uses and sound-word transformation model are all to adopt the training tool of setting to utilize popular corpus to train out, because Chinese vocabulary in popular corpus is all very abundant with expression, consider response user's time requirement, the length of phrase and the exponent number of model that adopt are all smaller, thereby have affected the accuracy to the sequence of candidate's Chinese sentence;

3) Chinese language model that identification translation module adopts and sound-word transformation model cannot centering sentence grammer, semanteme retrain, so can affect the accuracy to the sequence of candidate's Chinese sentence.

To sum up, the system that adopts prior art to provide is not high to the sequence accuracy of candidate's Chinese sentence, needs user by selecting module after loaded down with trivial details selection, just can obtain the Chinese sentence of needs, has reduced efficiency and Experience Degree.

Summary of the invention

In view of this, the invention provides a kind of optimization system of Chinese sentence input method, this system can be introduced additional features optimization candidate's Chinese sentence list, improves candidate's Chinese sentence list accuracy.

The present invention also provides a kind of optimization method of Chinese sentence input method, and the method can be introduced additional features optimization candidate's Chinese sentence list, improves candidate's Chinese sentence list accuracy.

For achieving the above object, technical scheme of the invention process is specifically achieved in that

An optimization system for Chinese sentence input method, this system comprises: input method engine modules, identification module, the module that reorders, candidate's Chinese sentence list display module, selection module and output module, wherein,

Input method engine modules, for resolving the pinyin string that obtains input, sends to identification module;

Identification module, for to the pinyin string receiving, identifies after translation, obtains multiple candidates' Chinese sentence, sends to the module that reorders;

Module reorders, be used for receiving multiple candidate's Chinese sentences, to each Chinese sentence, calculate the proper vector of reaction speech habits, syntax and semantics according to the resource file generating, carry out inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, according to the scoring height of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, send to candidate's Chinese sentence list display module;

Candidate's Chinese sentence list display module, for the list of show candidate Chinese sentence, by selecting module controls to select to obtain a Chinese sentence, exports by output module;

Select module, meet for of selecting candidate's Chinese sentence list display module the Chinese sentence that user requires most.

Described identification module, also for after obtaining multiple candidates' Chinese sentence, sends to the module that reorders before the Chinese sentence list of candidate to be sorted.

Described system also comprises: resource file storehouse and feature weight module is provided, wherein,

Resource file storehouse, for generating sound-word conversion table, mutual information table, Chinese language model and part-of-speech tagging language model, offers the module that reorders;

Feature weight module is provided, for generating feature weight, offers the module that reorders;

Module reorders, the Chinese sentence that will sort for each, also for obtaining based system eigenwert according to former sequence, calculate the mutual information of candidate's sentence according to mutual information table, obtain word word tone transition probability according to sound-word conversion table, reverse word word tone transition probability, word tone transition probability and reverse word tone transition probability, calculate the language model scoring of candidate's sentence according to Chinese language model, calculate the part-of-speech tagging language model scoring of candidate's sentence according to part-of-speech tagging language model, obtain candidate's sentence length according to the phrase number of Chinese sentence, and then obtained proper vector and feature weight are asked to inner product, obtain the scoring of each Chinese sentence, according to the scoring height of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, send to candidate's Chinese sentence list display module.

An optimization method for Chinese sentence input method, the method comprises the required resource file of generation calculated characteristics vector, training characteristics weight utilizes the proper vector of feature weight and expression candidate sentence to mark to candidate's sentence, is specially:

The pinyin string of user's input is identified after translation, obtained multiple candidates' Chinese sentence;

To each Chinese sentence, calculate the proper vector of reaction speech habits, syntax and semantics according to the resource file generating, carry out inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, according to scoring height, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, offer user;

From the list of candidate's Chinese sentence, select to obtain a Chinese sentence, output according to user's selection.

Described resource file comprises: one or more combinations in sound-word conversion table, mutual information table, Chinese language model and part-of-speech tagging language model,

Described resource file and described feature weight adopt with the Chinese language resource distributing and obtain.

The component of described proper vector is the combination of following one or more features: based system feature, mutual information, word word tone transition probability, reverse word word tone transition probability, word tone transition probability, reverse word tone transition probability, Chinese language model, part-of-speech tagging language model and candidate's sentence length.Wherein,

Obtain based system eigenwert according to the inverse of former sequence, calculate the mutual information of candidate's sentence according to mutual information table, obtain word word tone transition probability according to sound-word conversion table, reverse word word tone transition probability, word tone transition probability and reverse word tone transition probability, calculate the language model scoring of candidate's sentence according to Chinese language model, calculate the part-of-speech tagging language model scoring of candidate's sentence according to part-of-speech tagging language model, obtain candidate's sentence length according to the phrase number of Chinese sentence, and then obtained proper vector and feature weight are asked to inner product, obtain the scoring of each Chinese sentence, according to the scoring height of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence.

As seen from the above technical solution, the present invention is obtaining after multiple candidates' Chinese sentence, calculate for each Chinese sentence after the proper vector of reaction speech habits, syntax and semantics, carry out after inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, after multiple candidate's Chinese sentences being sorted according to the scoring of each Chinese sentence, show by candidate's Chinese sentence list display module.Because the candidate's Chinese sentence tab sequential that adopts sort method provided by the invention to obtain is that according to more meeting, user language is accustomed to, syntactical and semantical feature carries out tactic, thereby optimize candidate's Chinese sentence list, improve candidate's Chinese sentence list accuracy.

Brief description of the drawings

Fig. 1 is the system schematic of the Chinese sentence input method of prior art;

Fig. 2 is the optimization system schematic diagram of Chinese sentence input method provided by the invention;

Fig. 3 is the optimization system embodiment schematic diagram of Chinese sentence input method provided by the invention;

Fig. 4 is the optimization method process flow diagram of Chinese sentence input method provided by the invention.

Embodiment

For making object of the present invention, technical scheme and advantage clearer, referring to the accompanying drawing embodiment that develops simultaneously, the present invention is described in further detail.

Can find out from background technology, candidate's Chinese sentence tab sequential accuracy of the candidate's Chinese sentence list display module output in Chinese sentence input method system is not high, make user select first hit rate lower, increase number of times former that user selects because: carrying out in the alignment processes of candidate's Chinese sentence list, only sort according to the Chinese language model arranging, and do not have the speech habits of centering sentence, syntactical and semantical feature to consider.

Therefore, the present invention is directed to Chinese speech habits, syntactical and semantical feature, adopt with the Chinese language resource distributing and generate resource file training characteristics weight, obtaining after multiple candidates' Chinese sentence, for each Chinese sentence, calculate according to resource file after the proper vector of reaction speech habits, syntax and semantics, carry out after inner product operation with trained feature weight, obtain the scoring of Chinese sentence, after multiple candidate's Chinese sentences being sorted according to the scoring of each candidate's Chinese sentence, show by candidate's Chinese sentence list display module.Because the candidate's Chinese sentence tab sequential that adopts sort method provided by the invention to obtain is that according to more meeting, user language is accustomed to, syntactical and semantical feature carries out tactic, thereby optimize candidate's Chinese sentence list, improve candidate's Chinese sentence list accuracy.

In this embodiment, to each candidate's Chinese sentence, the combination that the component of the proper vector calculating is following one or more eigenwerts: based system eigenwert, word word tone transition probability, reverse word word tone transition probability, word tone transition probability, reverse word tone transition probability, mutual information, Chinese language model, part-of-speech tagging language model and candidate's sentence length.

Wherein, mutual information is for weighing long-distance dependence and the semantic consistency of candidate's sentence, word word tone transition probability converts the possibility of its pinyin string to for weighing the Chinese character string of word, reverse word word tone transition probability is for weighing the possibility that converts this word Chinese character string from the pinyin string of word to, word tone transition probability is changed the possibility of its phonetic for weighing the Chinese character of word, reverse word tone transition probability is converted to the possibility of its Chinese character for weighing the phonetic of word, Chinese language model is for weighing the fluency of candidate's Chinese sentence, part-of-speech tagging language model is used for weighing the grammatical degree of candidate's Chinese sentence.

Reverse word word tone transition probability, word tone transition probability and reverse word tone transition probability have reacted the accuracy of changing between the phonetic of Chinese sentence and Chinese character, can obtain according to sound-word conversion table.Mutual information, Chinese language model, part-of-speech tagging language model and candidate's sentence lengths table understand the syntax and semantics custom of Chinese sentence, mutual information can calculate according to the mutual information table generating, Chinese language model, part-of-speech tagging language model need respectively Chinese language model and obtain with the part-of-speech tagging language model of Part of Speech Tagging language material training, based system feature can adopt the inverse of candidate's Chinese sentence sequence number indirectly to obtain, thereby does not need resource file.

Word word tone transition probability, reverse word word tone transition probability, word tone transition probability, reverse word tone transition probability, mutual information and candidate's sentence length need to carry out participle to candidate's Chinese sentence, and part-of-speech tagging language model needs candidate's Chinese sentence to carry out participle and part-of-speech tagging.

Fig. 2 is the optimization system of Chinese sentence input method provided by the invention, comprising: input method engine modules, identification module, the module that reorders, candidate's Chinese sentence list display module, selection module and output module, wherein,

Input method engine modules, resolves the pinyin string that obtains user's input for the input method according to setting, send to identification module;

Identification module, for the pinyin string to the input of input method engine modules, identifies after translation, obtains multiple candidates' Chinese sentence, sends to the module that reorders;

Module reorders, the multiple candidate's Chinese sentences that send for receiving identification module, for each Chinese sentence, calculate the proper vector of reaction speech habits, syntax and semantics according to resource file, carry out inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, according to the scoring of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, send to candidate's Chinese sentence list display module;

Candidate's Chinese sentence list display module, for the list of show candidate Chinese sentence, by selecting module controls to select to obtain a Chinese sentence, exports by output module;

Select module, meet for of selecting candidate's Chinese sentence list display module the Chinese sentence that user requires most.

In the present invention, identification module also can adopt existing identification translation module, obtaining after multiple candidates' Chinese sentence, according to certain strategy, such as according to language model scoring, sort, obtain corresponding candidate's Chinese sentence list, send to the module that reorders, by the module that reorders introduce additional features to candidate Chinese sentence list carry out reordering again.Like this, the present invention has increased the module that reorders in the system-based shown in Fig. 1, according to the speech habits of Chinese, syntax and semantics, the Chinese sentence list of candidate is resequenced again.

In the present invention, the module that reorders needs use characteristic weight and resource file, as shown in Figure 3, resource file storehouse, provide the module of the feature weight order module of attaching most importance to that resource file and feature weight are provided, the resource file in resource file storehouse and the feature weight in feature weight module is provided is to be provided by the Chinese language resources bank distributing together.Particularly,

Chinese language resources bank, be used for according to Chinese language resource, such as the Chinese sentence adopting in the Peoples Daily, generate for generating the first language material of source file and the second language material of generating feature weight, offer respectively resource file storehouse and feature weight module is provided;

Resource file storehouse, for generate respectively sound-word conversion table resource file, mutual information table resource file, Chinese language model resource file and part-of-speech tagging language model resource file according to the first language material, offers the module that reorders;

Feature weight module is provided, for according to the second language material generating feature weight, offers the module that reorders;

Module reorders, the Chinese sentence that specifically will sort for each, obtain based system eigenwert according to former sequence, calculate the mutual information of candidate's sentence according to mutual information table, obtain word word tone transition probability according to sound-word conversion table, reverse word word tone transition probability, word tone transition probability and reverse word tone transition probability, calculate the language model scoring to candidate's sentence according to Chinese language model, calculate the part-of-speech tagging language model scoring to candidate's sentence according to part-of-speech tagging language model, obtain candidate's sentence length according to the phrase number of Chinese sentence, and then obtained proper vector and feature weight are asked to inner product, just obtain the comprehensive grading of this Chinese sentence, according to the height of scoring, multiple candidate's Chinese sentences are sorted, just can obtain final candidate's Chinese sentence list.

In this embodiment, the component of the feature obtaining can be the combination of above-mentioned one or more features, no longer limits here.

In this embodiment, specifically carrying out the reordering in process of the whole sentence of Chinese, need not comprise Chinese language resources bank, in resource file storehouse, generate but utilize the resource file obtaining according to Chinese language resource, and providing feature weight module to preserve the feature weight of training.

Fig. 4 is the optimization method process flow diagram of Chinese sentence input method provided by the invention, adopts in advance with the Chinese language resource training characteristics weight distributing and for calculating the resource file of reaction Chinese language custom, syntactical and semantical feature, and the method also comprises:

Step 401, Chinese sentence input method system receive the pinyin string of user's input;

Step 402, Chinese sentence input method system are identified the pinyin string of user's input after translation, obtain multiple candidates' Chinese sentence;

Step 403, Chinese sentence input method system are for each Chinese sentence, calculate the proper vector of reaction speech habits, syntax and semantics according to the resource file of preparing, carry out inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, according to scoring height, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, offer user;

Step 404, Chinese sentence input method system select to obtain a Chinese sentence, output according to user's selection from the list of candidate's Chinese sentence.

Below in conjunction with specific embodiment to how obtaining reacting the speech habits of Chinese, the source file of syntax and semantics is elaborated.

Prepare the first language material

The first language material is that the phonetic-Chinese character sentence obtaining according to Chinese language resource is right, for generating sound-word conversion table, mutual information table, Chinese language model and part-of-speech tagging language model.

The process of preparing the first language material is:

First step, from Chinese language resource, such as randomly drawing Chinese sentence in the Peoples Daily, a line storage, as document a;

Second step, by the Chinese sentence phonetic notation in document a, deposits a line of phonetic in document b, and in document b, in every a line phonetic and document a, the sentence of corresponding row is corresponding one by one;

Third step is randomly drawed approximately 1000 sentence strings from document b, and an a line forms document bb, then from document a and document b, deletes the sentence appearing in document bb;

The 4th step, carries out word segmentation to the sentence of deleting in the document a of sentence in document bb, adopts space to separate;

The 5th step, according to the participle of document a of deleting sentence in document bb, carry out corresponding participle to deleting corresponding pinyin string in the document b of sentence in document bb, the pinyin order of word in word is connected together, the pinyin string of different terms separates with space, obtains final document a and final document b.

The explanation of giving one example

Delete the sentence in the document a of sentence in document bb: she is a beautiful little girl.

Delete in the document b of sentence in document bb pinyin string that should sentence: ta shi yigemei ' li de xiao gu ' niang.

Generate sound-word conversion table resource file according to the first language material

Detailed process is:

First step, order reads the sentence in final document a and the final document b of the first language material, and composition sentence is right;

Second step, sound-word conversion times of statistics sentence centering word and word in final document a, outcome record is in the first temporary file temp1 arranging;

Third step, continues to carry out first step and second step, until by the sentence in final document a and final document b to all handling;

The 4th step, calculate word word tone transition probability, reverse word word tone transition probability, word tone transition probability, reverse word tone transition probability according to the sound-word conversion times recording in the first temporary file temp1, and by result of calculation get after denary logarithm with word to together with storage, as sound-word conversion table resource file.

In this process,

In the first temporary file temp1, data layout is: ta||| she || | 15579; Ta ' men||| they || | 6891; Wherein, first is the pinyin string in final document b, and second is word corresponding in final document a, and the 3rd is corresponding sound-word conversion times;

The formula that calculates word word tone transition probability and reverse word word tone transition probability is:

$p\left(h_i\right|y_i)=\frac{N(h_i,y_i)}{N\left(y_i\right)}---\left(2\right)$

$p\left(y_i\right|h_i)=\frac{N(y_i,h_i)}{N\left(h_i\right)}---\left(3\right)$

Wherein, N (h _i, y _i) be the pinyin string y of word _iconvert Chinese character string h to _inumber of times; N (y _i, h _i) be the Chinese character string h of word _iconvert pinyin string y to _inumber of times; N (h _i) be the Chinese character string h of word _ioccurrence number; N (y _i) be the pinyin string y of word _ithe number of times occurring.

Calculate word tone transition probability, and the formula of reverse word tone transition probability is:

$\mathrm{lex}\left(h\right|y)=\underset{i=1}{\overset{l}{\mathrm{\Π}}}p\left(h_i\right|y_i)---\left(4\right)$

$\mathrm{lex}\left(y\right|h)=\underset{i=1}{\overset{l}{\mathrm{\Π}}}p\left(y_i\right|h_i)---\left(5\right)$

Wherein, y and h are respectively pinyin string and the Chinese character of word.P (h _i| y _i) be Chinese character h in word h _ipinyin string y _iconvert Chinese character h to _iprobability; P (y _i| h _i) be Chinese character h in word h _iconvert its pinyin string y to _iprobability.L is the number of Chinese character in this word h.

The form of sound-word conversion table resource file is: pinyin string y||| Chinese character string h|||log (p (h|y)) log (lex (h|y)) log (p (y|h)) log (lex (y|h))

A data example in sound-word conversion table: ta ' men||| they || |-0.9586-1.1549-2.7959-3.3979.

Generate mutual information table resource file according to the first language material

Detailed process is:

First step, reads a sentence in the final document a of the first language material, and in statistics sentence, word and word are to the number of times occurring;

Second step, repeat first step until in final document a all sentences all processed complete;

Third step, calculates probability and the right co-occurrence probabilities of word of all words;

The 4th step, obtains the right mutual information of word according to the probability of word and the right co-occurrence probabilities of word, preserves as mutual information table resource file.

In this process,

The formula that calculates Word probability is:

$p\left(w\right)=\frac{N\left(w\right)}{N}---\left(6\right)$

Wherein, the number of times that N (w) occurs for word w; N is the sum of word in language material.

Calculate word to (w _i, w _j) the formula of co-occurrence probabilities be:

$p\left(w_i\right|w_j)=\frac{N(w_i,w_j)}{\underset{r}{\mathrm{\Σ}}N(w_r,w_j)}---\left(7\right)$

$p\left(w_j\right|w_i)=\frac{N(w_i,w_j)}{\underset{r}{\mathrm{\Σ}}N(w_i,w_r)}---\left(8\right)$

Wherein, N (w _i, w _j) be that word is to (w _i, w _j) number of times of co-occurrence; N (w _r, w _j) be that word is to (w _r, w _j) number of times of co-occurrence; N (w _i, w _r) be that word is to (w _i, w _r) number of times of co-occurrence.

Calculate word to (w _i, w _j) formula of mutual information is:

$I(w_i,w_j)=\log \frac{p(w_i,w_j)}{p\left(w_i\right)p\left(w_j\right)}=\log \frac{p\left(w_i\right|w_j)}{p\left(w_i\right)}=\log \frac{p\left(w_j\right|w_i)}{p\left(w_j\right)}---\left(9\right)$

Wherein, p (w _i| w _j) for there is word w _jand there is word w _iprobability in same sentence; P (w _j| w _i) there is word w in sentence _iand there is word w simultaneously _jprobability; P (w _i| w _j) and p (w _j| w _i) computing method referring to formula (7) and formula (8); P (w _i) and p (w _j) be respectively word w _iwith word w _jthe probability occurring, computing method are referring to formula (6).

In mutual information table resource file, the form of data is: word w _i|| | word w _j|| | I (w _i, w _j)

A data example in mutual information table: she || | Miss || |-0.2825

Generate Chinese language model resource file according to the first language material

Process is, adopts SRILM to train 4 rank Chinese language models to the final document a of the first language material, and wherein, SRILM is existing language model training tool.

The example of a Chinese language model resource file :-2.712682 she be

-2.712343 she want

Generate part-of-speech tagging language model resource file according to the first language material

Process is:

First step, carries out part-of-speech tagging to the sentence in the final document a of the first language material, produces after part-of-speech tagging string, is kept in the second temporary file temp2 with the form an of a line;

Second step, taking the second temporary file temp2 as input, adopts SRILM to train 7 rank part-of-speech tagging language model source files.

Data example in the second temporary file temp2, such as: " I am a student to the sentence in corresponding final document a.", in the second temporary file temp2, corresponding part-of-speech tagging string is: " r vm q n w ".

The example of data in a part-of-speech tagging language model resource file :-0.325047 r v

Prepare the second language material

The second language material is the language material for training characteristics weight, and it is obtained by Chinese language resource, requires each pinyin string in this language material to have 10 candidates' Chinese sentence.

The process of preparing the second language material is:

First step, the each sentence in document bb is adopted to the form input of the corresponding sentence string of pinyin string, and choose front 10 candidate's sentences, be kept in the 3rd temporary file temp3;

Second step, to the sentence participle in the 3rd temporary file temp3, and marks part of speech, stores in document c, and document c is the second language material.

An example of data layout in the 3rd transient document temp3:

<corpus id＝“2”>

<bead id＝“1”>

<pinyin num＝“20”>

ta shi yi ge mei’li de xiao gu’niang

</pinyin>

<sent id＝“1”>

She is a beautiful little girl

</sent>

<sent id＝“2”>

He is a beautiful little girl

</sent>

<sent id＝“3”>

He is a bonnily little girl

</sent>

<sent id＝“4”>

It is one and beautiful laughs at Miss

</sent>

……

</bead>

<bead id＝“2”>

……

</bead>

……

</corpus>

Data layout example in document c:

<corpus id＝“2”>

<bead id＝“1”>

<pinyin num＝“20”>

ta shi yi ge mei’li de xiao gu’niang

</pinyin>

<sent id＝“1”>

She/rr is /vshi mono-/mq/q beauty/an/dec is little/Miss an/n

</sent>

<sent id＝“2”>

He/rr is /vshi mono-/mq/q beauty/an/dec is little/Miss an/n

</sent>

<sent id＝“3”>

He/rr is /vshi mono-/mq/q beauty/an ground/di is little/Miss an/n

</sent>

<sent id＝“4”>

It/rr is /vshi mono-/mq/laugh at/Miss v/n of q beauty/an/dec

</sent>

……

</bead>

<bead id＝“2”>

……

</bead>

……

</corpus>

According to the second language material training characteristics weight

From the second language material, choose m pinyin string, { y ₁, y ₂..., y _m, each pinyin string y _i∈ { y ₁, y ₂..., y _mthere are n candidate Chinese character string, { a h _i1, h _i2..., h _in.

The arthmetic statement of training characteristics weight is as follows:

Input: a positive slack variable τ, τ=0.0001 in the present embodiment.Iterations t, initial value is t=0.Feature weight initial value is: w ⁰=[0,0 ..., 0]

Output: feature weight

The algorithm of calculated characteristics weight is:

repeat{

For (m) do{ of pinyin string subscript i=1 to

For (candidate's sentence subscript j=1 to n-1) do{

u _j＝[0，0，…，0]

For (n) do{ of candidate's sentence subscript 1=j+1 to

$\mathrm{if}(w^t\&CenterDot;x_{\mathrm{ij}}<w^t\&CenterDot;x_{\mathrm{il}}+(\frac{1}{j}-\frac{1}{l})\mathrm{\&tau;})\{$

$u_j=u_j+(\frac{1}{j}-\frac{1}{l}){(x_{\mathrm{ij}}-x_{\mathrm{il}})}^T---\left(10\right)$

}

}

}

$w^{t+1}=w^t+\underset{j}{\mathrm{\&Sigma;}}{u}_j$

(11)

}

Till until feature weight w is not modified

Illustrate: (x in above-mentioned algorithm _ij-x _il) ^tfor column vector (x _ij-x _il) transposition.

The concrete calculating of the proper vector to each candidate's Chinese sentence

Calculate component corresponding to based system eigenwert:

If the sequence number of the former sequence of candidate Chinese sentence is followed successively by: 1,2,3 ... natural number.Component corresponding to this feature is the inverse of candidate's Chinese sentence sequence number.

Calculate component corresponding to mutual information:

If h _ibe i ^thindividual candidate's sentence.H _i={ h _i1, h _i2..., h _in, h _ij∈ { h _i1, h _i2..., h _inbe candidate's sentence h _iin a word.The component that this candidate sentence mutual information is corresponding is:

$\mathrm{MI}\left(h_i\right)=\underset{j=1}{\overset{n-1}{\mathrm{\&Sigma;}}}\underset{l=j+1}{\overset{n}{\mathrm{\&Sigma;}}}I(h_{\mathrm{ij}},h_{\mathrm{il}})---\left(12\right)$

Wherein, I (h _ij, h _il) value can from mutual information table resource file, search, find and return to its value, otherwise return to 0.

Word word tone transition probability and the component corresponding to reverse word word tone transition probability of calculated candidate Chinese sentence:

If h _ibe i ^thindividual candidate's sentence.H _i={ h _i1, h _i2..., h _in, h _il∈ { h _i1, h _i2..., h _inbe candidate's sentence h _iin a word.If y is the pinyin string of candidate's sentence, y={y ₁, y ₂..., y _n, y _j∈ { y ₁, y ₂..., y _nbe with candidate's sentence in word h _ijcorresponding pinyin string, n is the number of word in sentence.Candidate's Chinese sentence h _iword word tone transition probability and component corresponding to reverse word word tone transition probability be respectively:

$\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\log p\left(y_j\right|h_{\mathrm{ij}})---\left(13\right)$

$\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\log p\left(h_{\mathrm{ij}}\right|y_j)---\left(14\right)$

Wherein, logp (y _j| h _ij) and logp (h _ij| y _j) value from sound-word conversion table resource file, search, if search less than, return to 0.

Word tone transition probability and the component corresponding to reverse word tone transition probability of calculated candidate Chinese sentence:

If h _ibe i ^thindividual candidate's sentence.H _i={ h _i1, h _i2..., h _in, h _ij∈ { h _i1, h _i2..., h _inbe candidate's sentence h _iin a word.If y is the pinyin string of candidate's sentence, y={y ₁, y ₂..., y _n, y _j∈ { y ₁, y ₂..., y _nbe with candidate's sentence in word h _ijcorresponding pinyin string, n is the number of phrase in sentence.Word tone transition probability and component corresponding to reverse word tone transition probability are:

$\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\log \mathrm{lex}\left(y_j\right|h_{\mathrm{ij}})---\left(15\right)$

$\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\log \mathrm{lex}\left(h_{\mathrm{ij}}\right|y_j)---\left(16\right)$

Wherein, loglex (h _ij| y _j) and loglex (y _j| h _ij) value from sound-word conversion table resource file, search, if can not find, return to 0.

Component corresponding to calculated candidate Chinese sentence language model:

If h _ibe i ^thindividual candidate's sentence.H _i={ h _i1, h _i2..., h _in, h _ij∈ { h _i1, h _i2..., h _inbe candidate's sentence h _iin a word.The language model using in the present embodiment is 4 rank language models, and component corresponding to candidate's Chinese sentence language model is:

P _lm＝p(h _i1)*p(h _i2|h _i1)*p(h _i3|h _i1h _i2)*p(h _i4|h _i1h _i2h _i3)

*p(h _i5|h _i2h _i3h _i4)*…*p(h _in|h _in-3h _in-2h _in-1) (17)

Wherein, each probable value is looked into and is got from Chinese language model resource file.

The explanation of giving one example, Chinese candidate sentence: I am a student;

Language model eigenwert is: P _lm=p (I) * p (be | I) * p (one | I am) * p (individual I be one) * p (student | be one).

The component corresponding to part-of-speech tagging language model of calculated candidate Chinese sentence:

If tag _ibe i ^thindividual candidate's sentence h _imark sequence.Tag _i={ tg _i1, tg _i2..., tg _in, tg _ij∈ { tg _i1, tg _i2..., tg _inbe candidate's sentence h _iin the mark of j word.The part-of-speech tagging language model using in the present embodiment is the language model on 7 rank, the part-of-speech tagging language model pair of candidate's Chinese sentence

P _lm-pos＝p(tg _i1)*p(tg _i2|tg _i1)*p(tg _i3|tg _i1tg _i2)*p(tg _i4|tg _i1tg _i2tg _i3)

*p(tg _i5|tg _i1tg _i2tg _i3tg _i4)*p(tg _i6|tg _i1tg _i2tg _i3tg _i4tg _i5)

*p(tg _i7|tg _i1tg _i2tg _i3tg _i4tg _i5tg _i6)*p(tg _i8|tg _i2tg _i3tg _i4tg _i5tg _i6tg _i7)

The component of answering is: * ... * p (tg _in| tg _in-6tg _in-5tg _in-4tg _in--3tg _in-2tg _in-1) (18)

Wherein, each probable value is looked into and is got from part-of-speech tagging language model resource file.

Lift an object lesson explanation, Chinese candidate sentence: I am a student;

Mark sequence: r v m q n

The component that part-of-speech tagging language model is corresponding is:

P _lm-pos＝p(r)*p(v|r)*p(m|rv)*p(q|rvm)*p(n|rvmq)

Component corresponding to calculated candidate Chinese sentence length:

It is length characteristic value that this example is got phrase number in candidate's Chinese sentence.If h _ibe i ^thindividual candidate's Chinese sentence.H _i={ h _i1, h _i2..., h _in, h _ij∈ { h _i1, h _i2..., h _inbe candidate's Chinese sentence h _iin a word.Candidate's Chinese sentence h _icomponent corresponding to length be n.

Below for three examples that adopted the module that reorders to reorder.

Example 1:

The pinyin string of user's input is: haizidalelianghujiangyou

Output candidate is: child is large, and two families will have

Child has bought two kettle soy sauce

Be output as with after reordering: child has bought two kettle soy sauce

Child is large, and two families will have

Example 2:

Input Pinyin string: jinlaijiangyutebieduo

Output candidate is: come in many especially

Recently rainfall is many especially

Be output as with after reordering: rainfall is recently many especially

Coming in will be in many especially

Example 3:

Input Pinyin string: tashigeguniang

Output candidate is: he is a Miss

She is a Miss

Be output as with after reordering: she is a Miss

He is a Miss

To sum up, the proper vector of the composition such as Chinese sentence basis such as mutual information, multiple sound-word transition probability, Chinese language model, part-of-speech tagging language model and sentence length of the present invention to multiple candidates is carried out sequence again from grammer, semantic aspect, obtained good effect:

1) it is more reasonable that it can make the sequence of candidate's Chinese sentence list, increased the initial hit rate that user selects, and reduces the number of times that user selects, and accelerated the speed that user selects;

2) the word word tone transition probability in proper vector, reverse word word tone transition probability, word tone transition probability, reverse word tone transition probability, Chinese language model, part-of-speech tagging language model and/or mutual information can obtain according to Chinese language resource, utilize these features to reorder to candidate's Chinese sentence, can make the candidate's Chinese sentence that more meets user language custom, syntax and semantics in the list of candidate's Chinese sentence come more forward position;

3) mutual information in proper vector has reacted in whole sentence complementary feature between word, by reordering, can make the candidate's Chinese sentence that more meets user language custom come more forward position;

4) Chinese language model in proper vector and part-of-speech tagging language model can be weighed candidate Chinese sentence and meet the degree of Chinese grammer, and the sentence sequence that makes more to meet in candidate's sentence Chinese grammar request is more forward.

More than lift preferred embodiment; the object, technical solutions and advantages of the present invention are further described; institute is understood that; the foregoing is only preferred embodiment of the present invention; not in order to limit the present invention; within the spirit and principles in the present invention all, any amendment of doing, be equal to and replace and improvement etc., within all should being included in protection scope of the present invention.

Claims (4)

1. an optimization system for Chinese sentence input method, is characterized in that, this system comprises: input method engine modules, identification module, the module that reorders, candidate's Chinese sentence list display module, selection module and output module, wherein,

Input method engine modules, for resolving the pinyin string that obtains input, sends to identification module;

Identification module, for to the pinyin string receiving, identifies after translation, obtains multiple candidates' Chinese sentence, sends to the module that reorders;

Module reorders, be used for receiving multiple candidate's Chinese sentences, to each Chinese sentence, calculate the proper vector of reaction speech habits, syntax and semantics according to the resource file generating, carry out inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, according to the scoring height of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, send to candidate's Chinese sentence list display module;

Candidate's Chinese sentence list display module, for the list of show candidate Chinese sentence, by selecting module controls to select to obtain a Chinese sentence, exports by output module;

Select module, meet for of selecting candidate's Chinese sentence list display module the Chinese sentence that user requires most;

Described system also comprises: resource file storehouse and feature weight module is provided, wherein,

Resource file storehouse, for generating sound-word conversion table, mutual information table, Chinese language model and part-of-speech tagging language model, offers the module that reorders;

Feature weight module is provided, for generating feature weight, offers the module that reorders;

Module reorders, the Chinese sentence that will sort for each, also for obtaining based system eigenwert according to former sequence, calculate the mutual information of candidate's sentence according to mutual information table, calculate word word tone transition probability according to sound-word conversion table, reverse word word tone transition probability, word tone transition probability and reverse word tone transition probability, calculate the language model scoring of candidate's sentence according to Chinese language model, calculate the part-of-speech tagging language model scoring of candidate's sentence according to part-of-speech tagging language model, obtain candidate's sentence length according to the phrase number of Chinese sentence, and then obtained proper vector and feature weight are asked to inner product, obtain the scoring of each Chinese sentence, according to the scoring height of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, send to candidate's Chinese sentence list display module,

The inverse of the sequence number of the former sequence that component corresponding to described based system eigenwert is described Chinese sentence;

The described formula that calculates word word tone transition probability and reverse word word tone transition probability is: $p\left(h_i\right|y_i)=\frac{N(h_i,y_i)}{N\left(y_i\right)},$ $p\left(y_i\right|h_i)=\frac{N(y_i,h_i)}{N\left(h_i\right)},$ Wherein, N (h _i, y _i) be the pinyin string y of word _iconvert Chinese character string h to _inumber of times; N (y _i, h _i) be the Chinese character string h of word _iconvert pinyin string y to _inumber of times; N (h _i) be the Chinese character string h of word _ioccurrence number; N (y _i) be the pinyin string y of word _ithe number of times occurring;

The described formula that calculates word tone transition probability and reverse word tone transition probability is: $\mathrm{lex}\left(h\right|y)=\underset{i=1}{\overset{l}{\mathrm{\&Pi;}}}p\left(h_i\right|y_i),$ $\mathrm{lex}\left(y\right|h)=\underset{i=1}{\overset{l}{\mathrm{\&Pi;}}}p\left(y_i\right|h_i),$ Wherein, y and h are respectively pinyin string and the Chinese character of word; P (h _i| y _i) be Chinese character h in word h _ipinyin string y _iconvert Chinese character h to _iprobability; P (y _i| h _i) be Chinese character h in word h _iconvert its pinyin string y to _iprobability; L is the number of Chinese character in this word h.

2. the system as claimed in claim 1, is characterized in that, described identification module, also for after obtaining multiple candidates' Chinese sentence, sends to the module that reorders before the Chinese sentence list of candidate to be sorted.

3. the optimization method of a Chinese sentence input method, it is characterized in that, the method comprises the required resource file of generation calculated characteristics vector, training characteristics weight, utilize the proper vector of feature weight and expression candidate sentence to mark to candidate's sentence, be specially:

The pinyin string of user's input is identified after translation, obtained multiple candidates' Chinese sentence;

To each Chinese sentence, calculate the proper vector of reaction speech habits, syntax and semantics according to the resource file generating, carry out inner product operation with trained feature weight, obtain the scoring of each Chinese sentence, according to scoring height, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence, offer user;

From the list of candidate's Chinese sentence, select to obtain a Chinese sentence, output according to user's selection;

The component of described proper vector is the combination of following one or more features: based system feature, mutual information, word word tone transition probability, reverse word word tone transition probability, word tone transition probability, reverse word tone transition probability, Chinese language model, part-of-speech tagging language model and candidate's sentence length, wherein

Obtain based system eigenwert according to the inverse of former sequence, calculate the mutual information of candidate's sentence according to mutual information table, calculate word word tone transition probability according to sound-word conversion table, reverse word word tone transition probability, word tone transition probability and reverse word tone transition probability, calculate the language model scoring of candidate's sentence according to Chinese language model, calculate the part-of-speech tagging language model scoring of candidate's sentence according to part-of-speech tagging language model, obtain candidate's sentence length according to the phrase number of Chinese sentence, and then obtained proper vector and feature weight are asked to inner product, obtain the scoring of each Chinese sentence, according to the scoring height of each Chinese sentence, multiple candidate's Chinese sentences are sorted, obtain the list of candidate's Chinese sentence,

Component corresponding to described based system eigenwert be described Chinese sentence sequence number former sequence inverse;

The described formula that calculates word word tone transition probability and reverse word word tone transition probability is: $p\left(h_i\right|y_i)=\frac{N(h_i,y_i)}{N\left(y_i\right)},$ $p\left(y_i\right|h_i)=\frac{N(y_i,h_i)}{N\left(h_i\right)},$ Wherein, N (h _i, y _i) convert Chinese character string h to for the pinyin string yi of word _inumber of times; N (y _i, h _i) be the Chinese character string h of word _iconvert pinyin string y to _inumber of times; N (h _i) be the Chinese character string h of word _ioccurrence number; N (y _i) be the number of times of the pinyin string yi appearance of word;

The described formula that calculates word tone transition probability and reverse word tone transition probability is: $\mathrm{lex}\left(h\right|y)=\underset{i=1}{\overset{l}{\mathrm{\&Pi;}}}p\left(h_i\right|y_i),$ $\mathrm{lex}\left(y\right|h)=\underset{i=1}{\overset{l}{\mathrm{\&Pi;}}}p\left(y_i\right|h_i),$ Wherein, y and h are respectively pinyin string and the Chinese character of word; P (h _i| y _i) be Chinese character h in word h _ipinyin string y _iconvert Chinese character h to _iprobability; P (y _i| h _i) be Chinese character h in word h _iconvert its pinyin string y to _iprobability; L is the number of Chinese character in this word h.

4. method as claimed in claim 3, is characterized in that, described resource file comprises: one or more combinations in sound-word conversion table, mutual information table, Chinese language model and part-of-speech tagging language model,

Described resource file and described feature weight adopt with the Chinese language resource distributing and obtain.